NEAT

Name note: NEAT in this repository means Nash-Equilibrium Adaptive Training. It is unrelated to NeuroEvolution of Augmenting Topologies (NEAT; Stanley and Miikkulainen, 2002).

Nash-Equilibrium Adaptive Training

NEAT, short for Nash-Equilibrium Adaptive Training, is a Keras-first optimizer library for conflict-aware neural network optimization. It keeps the optimizer math explicit and testable in a NumPy reference engine, while exposing a small public API for real model training and an optional native CPU acceleration path.

A conflict-aware Keras optimizer with a versioned NumPy core

pip install "neat-optim[keras]" tensorflow

---

NEAT detects when the current gradient conflicts with the optimizer's opponent signal and reduces the conflicting component before updating. The correction is gated on a measured conflict ratio and clipped relative to the gradient norm.

• opponent source selection: momentum, previous_gradient, or gradient_ema
• correction warmup via correction_warmup_steps
• conflict gating via conflict_threshold
• adaptive alpha via adaptive_alpha
• diagonal second-moment preconditioning via adaptive_preconditioning
• Lion-style sign updates via update_mode="lion"
• gradient centralization via gradient_centralization
• Nesterov-style momentum via nesterov
• Lookahead slow-weight synchronization via lookahead_k
• per-run optimizer diagnostics via diagnostic_snapshot() or epoch-level TensorBoard logging with NEATDiagnosticsCallback

The update math lives in a versioned NumPy reference engine, fully independent of Keras. The Keras optimizer is a thin adapter on top. An optional C++ CPU extension drops in behind the same semantics.

• Public training API: Keras optimizer subclass
• Reference core: NumPy
• Native acceleration: optional CPU-only C++ extension
• Tested platforms: Linux and macOS
• Supported Python versions: 3.10 to 3.13
• Supported Keras setup: install keras plus a backend runtime such as TensorFlow. The optimizer is written against Keras 3 APIs; TensorFlow is the currently exercised integration backend in this repo.
• Optional PyTorch adapter: neat_optim.TorchNEAT, used by the modern reproducible benchmark suite.

The v3 engineering work is additive. nce_spec_v2 remains the default and its behavior is frozen; v3 research options are not promoted without an explicit author decision and full cross-engine parity coverage.

To select the additive specification, pass spec_version="nce_spec_v3". Default-off v3 controls cover parameter-local conflict granularity, an L2 trust-region correction bound, scheduled momentum/gradient-EMA opponent ensembles, correction/preconditioner composition order, and deterministic reference reductions. Existing constructors that omit spec_version continue to use v2 behavior.

For large PyTorch models, TorchNEAT(..., diagnostic_interval=10) samples diagnostics every ten steps. This reduces accelerator synchronization cost and does not alter optimizer updates.

Modern Benchmarks

The repository includes runnable comparisons for CIFAR-10/100 and ImageNet folders with ResNet/ViT/DeiT, GLUE SST-2 and Alpaca small-LLM fine-tuning, MuJoCo control, MNIST diffusion, and large-batch stability. Each run records convergence, loss variance, wall time, environment metadata, and optimizer diagnostics. See the benchmark guide. The dataset-free python -m benchmarks.torch_suite.compare_all --quick gate runs five seeds across NEAT, AdamW, SGD momentum, Lion, PCGrad, and CAGrad in their applicable synthetic settings.

What's in v0.1.0

First release. Everything listed here is shipping in 0.1.0. The public API is intentionally narrow — these are the pieces that have been validated end-to-end.

Core algorithm

The NEAT update rule (nce_spec_v1) introduces a Nash Conflict Elimination (NCE) correction term:

conflict = relu( −cos(g, o) ) # positive only when gradient opposes momentum
proj = proj_o(g) # project gradient onto the opponent direction
nce = −α · conflict · proj # subtract the conflicting component
u = g + nce # corrected gradient
m = β·m_prev + (1−β)·u # standard momentum
θ_next = (1 − lr·wd)·θ − lr·m # decoupled weight decay (default)

The correction is clipped: ‖nce‖ ≤ nce_clip_ratio · ‖g‖. If there is no conflict, nce = 0 and NEAT degrades to standard momentum SGD.

The full specification — including invariant proofs and a worked numerical example — is in docs/research/math-spec.md.

Shipped in this release

Component	Description
NEAT Keras optimizer	Drop-in model.compile() optimizer, fully serializable via get_config()
NumPy reference engine	Canonical implementation used for correctness validation and framework-free use
Functional API	neat_step(param, grad, state, config) — works without Keras or TensorFlow
Player-aware mode	Treats each per-example gradient as a separate player in a custom TF training loop
Native CPU extension	Optional C++ extension (pybind11) with automatic fallback to the reference engine
Model compaction	compact_dense_model() and search_compact_dense_model() for sparse model search
Diagnostics API	diagnostic_snapshot() to measure correction activity during and after training

Configurable in this release

Opponent signal — what counts as the conflict opponent

Value	Behaviour
"momentum" (default)	Uses the accumulated momentum vector
"gradient_ema"	Uses an exponential moving average of raw gradients
"previous_gradient"	Uses the previous raw gradient
"blended"	Weighted mix of momentum and gradient EMA

Correction controls — when and how much to correct

Parameter	Default	Effect
alpha	0.25	Correction strength — 0 disables correction entirely
nce_clip_ratio	1.0	Max correction size as fraction of gradient norm
correction_warmup_steps	0	Steps before correction activates
conflict_threshold	0.0	Minimum conflict ratio to apply any correction
nce_mode	"projection"	"projection", "cosine", or "off"

Adaptive features — optional scaling and preconditioning

Parameter	Default	Effect
adaptive_correction	False	Scale correction by opponent reliability and conflict EMA
adaptive_correction_decay	0.9	EMA decay for adaptive correction
adaptive_correction_min_scale	1.0	Lower clamp on adaptive scale
adaptive_correction_max_scale	3.0	Upper clamp on adaptive scale
adaptive_preconditioning	False	Adam-style second-moment preconditioner
second_moment_beta	0.999	Decay for second-moment EMA
bias_correction	False	Bias-correct momentum and second moment

Sparsity and weight decay

Parameter	Default	Effect
weight_decay	0.0	L2 weight decay coefficient
decouple_weight_decay	True	Apply decay before the gradient step (AdamW-style)
sparsity_l1	0.0	Soft L1 shrinkage after each step
prune_threshold	0.0	Hard-zero weights below this magnitude

Known limitations in v0.1.0

• Native C++ extension is CPU-only — no GPU kernel yet
• Player-aware mode requires TensorFlow as the Keras backend
• Benchmarks are on the sklearn digits dataset only — not ImageNet-scale

---

Installation

With TensorFlow (Linux / macOS):

pip install "neat-optim[keras]" tensorflow

With PyTorch (any platform, including Apple Silicon):

pip install "neat-optim[keras]" torch

With JAX:

pip install "neat-optim[keras]" jax

NumPy core only (no framework required):

pip install neat-optim

Local development:

python -m venv .venv &&source .venv/bin/activate
pip install -e ".[dev,keras]"&& pip install tensorflow

---

Quick Start

Standard Keras training

importkerasfromneat_optimimportNEATmodel=keras.Sequential(
[
keras.layers.Input((32,)),
keras.layers.Dense(64, activation="relu"),
keras.layers.Dense(10),
]
)
optimizer=NEAT(
learning_rate=1e-3,
alpha=0.25,
beta=0.9,
nce_mode="projection",
opponent_source="gradient_ema",
correction_warmup_steps=5,
adaptive_alpha=True,
adaptive_preconditioning=True,
)
model.compile(
optimizer=optimizer,
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=["accuracy"],
)
model.fit(x_train, y_train, epochs=20, validation_data=(x_val, y_val))
# See how much the correction engagedprint(model.optimizer.diagnostic_snapshot())

Custom training loop

importtensorflowastfimportkerasfromneat_optimimportNEAToptimizer=NEAT(learning_rate=1e-3, alpha=0.25)
loss_fn=keras.losses.SparseCategoricalCrossentropy(from_logits=True)
@tf.functiondeftrain_step(x, y):
withtf.GradientTape() astape:
logits=model(x, training=True)
loss=loss_fn(y, logits)
grads=tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
returnlossforx_batch, y_batchindataset:
loss=train_step(x_batch, y_batch)

Framework-free NumPy engine

importnumpyasnpfromneat_optimimportNEATConfigfromneat_optim.engine.functionalimportneat_stepfromneat_optim.stateimportArrayStateparam=np.array([1.0, -2.0, 0.5], dtype=np.float32)
grad=np.array([0.5, -0.25, 0.1], dtype=np.float32)
state=ArrayState.zeros_like(param)
config=NEATConfig(learning_rate=0.1, alpha=0.25, beta=0.9)
for_inrange(10):
result=neat_step(param, grad, state, config)
param=result.paramstate=result.stateprint(result.metrics)
# StepMetrics(grad_norm=0.559, conflict_ratio=0.012, nce_norm=0.001, ...)

Per-example player-aware training

importkerasfromneat_optimimportPlayerNEATConfigfromneat_optim.trainingimportcreate_player_states, player_train_stepmodel=keras.Sequential([
keras.layers.Input((32,)),
keras.layers.Dense(64, activation="relu"),
keras.layers.Dense(10),
])
loss_fn=keras.losses.SparseCategoricalCrossentropy(
from_logits=True, reduction="none"
)
states=create_player_states(model)
config=PlayerNEATConfig(
learning_rate=1e-2,
alpha=0.25,
sparsity_l1=1e-4,
prune_threshold=1e-3,
)
forx_batch, y_batchindataset:
result=player_train_step(model, x_batch, y_batch, loss_fn, states, config)
states=result.statesprint(f"loss={result.loss:.4f}")

---

Public API

• neat_optim.NEAT
• neat_optim.NEATDiagnosticsCallback
• neat_optim.NEATConfig
• neat_optim.PlayerNEATConfig
• neat_optim.ArrayState
• neat_optim.engine.functional.neat_step
• neat_optim.engine.multiplayer.neat_player_step
• neat_optim.training.player_train_step

Diagnostics

Call diagnostic_snapshot() after training to check whether the correction engaged:

snap=model.optimizer.diagnostic_snapshot()
# {# "mean_conflict_ratio": 0.021, # average conflict detected per step# "mean_correction_ratio": 0.008, # correction size / gradient size# "mean_update_alignment": 0.997, # cosine similarity of corrected vs raw gradient# "mean_opponent_norm": 0.043, # magnitude of the opponent signal# "correction_active_fraction": 0.74, # fraction of steps where correction fired# }

Interpreting results: If mean_conflict_ratio < 0.01 and correction_active_fraction < 0.1, the correction is essentially inactive on this task. NEAT will behave like standard momentum SGD and tuning alpha won't help. Tasks with noisy or multi-objective gradients tend to show much higher conflict ratios.

Optional research modes extend the same base update without changing defaults:

• adaptive_alpha=True makes the correction strength respond to conflict, gradient-noise, and alignment trends.
• adaptive_preconditioning=True adds a cheap diagonal second-moment preconditioner, similar in spirit to Fisher/Hessian-diagonal scaling.
• update_mode="lion" applies a sign-based final update after the NEAT correction.
• gradient_centralization=True, nesterov=True, and lookahead_k > 0 enable standard training stabilizers that are useful benchmark ablations.

This is Nash-inspired because the optimizer reacts to directional conflict between the current gradient and an opponent proxy. The exact behavior is defined in docs/research/math-spec.md.

Reset before each epoch to get per-epoch measurements:

forepochinrange(epochs):
model.optimizer.reset_diagnostics()
model.fit(x_train, y_train, epochs=1, verbose=0)
snap=model.optimizer.diagnostic_snapshot()
print(f"Epoch {epoch+1}: conflict={snap['mean_conflict_ratio']:.4f}, "f"active={snap['correction_active_fraction']:.2%}")

---

Benchmarks

The repository includes both lightweight local checks and larger benchmark harnesses:

• lint checks via ruff
• package build via python -m build
• docs build via mkdocs build --strict
• Keras integration tests
• reference/native parity tests
• a real Keras MLP benchmark against SGD, Adam, and AdamW
• a real Keras CNN benchmark on MNIST and Fashion-MNIST
• runnable benchmark harnesses for CIFAR-10 and GLUE SST-2
• benchmark diagnostics and sweep tooling for NEAT-specific ablations

Evaluated on the sklearn digits dataset, two-layer MLP, 20 epochs, 3 seeds:

Optimizer	Test Accuracy	Notes
SGD + momentum	97.04%	lr=0.01, momentum=0.9
Adam	96.85%	default hyperparameters
AdamW	96.85%	weight_decay=1e-4
NEAT	94.72%	α=0.25, β=0.9, gradient_ema opponent

On the current 20-epoch Keras digits benchmark, NEAT reaches 0.9472 mean test accuracy and trails SGD/Adam baselines. The attached diagnostics show why that is plausible: the mean correction ratio is only 0.00385 and the mean gradient/update alignment is 0.99991, so NEAT is behaving very close to its base update on this well-aligned task.

On a stronger short-transfer benchmark with a small CNN on MNIST and Fashion-MNIST over 3 seeds and 2 epochs, adaptive NEAT now reaches the best mean test accuracy on both datasets: 0.9861 vs 0.9856 for Adam on MNIST, and 0.8786 vs 0.8725 for Adam on Fashion-MNIST. That is better evidence than the digits-only benchmark, but it is still not a substitute for broader GPU-side benchmarks such as ImageNet or GLUE.

To reproduce the benchmark:

python benchmarks/run.py # reproduce the table above
python benchmarks/sweep_neat.py # hyperparameter sweep
python benchmarks/compare_optimizers.py # side-by-side comparison

Full methodology and results: docs/research/benchmarks.md.

---

Repository Layout

neat-optim/
├── src/neat_optim/
│ ├── engine/
│ │ ├── reference.py # canonical NumPy implementation (nce_spec_v1)
│ │ ├── functional.py # public neat_step() API with native fallback
│ │ ├── multiplayer.py # per-player stepping
│ │ └── native.py # native extension bridge
│ ├── training/
│ │ └── tensorflow_players.py # TF player-aware helpers
│ ├── config.py # NEATConfig, PlayerNEATConfig
│ ├── state.py # ArrayState, StepResult, StepMetrics
│ ├── keras_optimizer.py # Keras adapter
│ └── compaction.py # dense model pruning utilities
├── cpp/neat_core/ # optional C++ extension (pybind11)
├── tests/
│ ├── unit/ # algorithm correctness, config validation
│ ├── integration/ # Keras optimizer, player training, compaction
│ ├── property/ # randomized invariant checks
│ └── regression/ # reproducibility of documented examples
├── benchmarks/ # reproducible benchmark entrypoints
├── examples/ # minimal runnable scripts
└── docs/
├── research/ # math spec, algorithm, player-aware, benchmarks
└── contributing/ # architecture, development, release workflow

---

To run the short standard vision benchmark:

python benchmarks/vision_adaptive_neat_vs_baselines.py

To run the CIFAR-10 benchmark harness:

python benchmarks/cifar10_adaptive_neat_vs_baselines.py

To run the GLUE SST-2 benchmark harness:

python benchmarks/glue_sst2_adaptive_neat_vs_baselines.py

The CIFAR-10 and GLUE SST-2 harnesses are included so the repo can scale to stronger benchmark environments. They are runnable here, but full credible ImageNet- or broad-GLUE-style evidence still requires a stronger machine than this local CPU-only setup.

Development

ruff check .# lint
ruff format .# format
mypy src/neat_optim # type check
pytest tests/unit # fast, no Keras required
pytest # full suite
pytest --cov=neat_optim --cov-report=term-missing # with coverage
python -m build # build wheel + sdist
mkdocs build --strict # docs build

See CONTRIBUTING.md for the full workflow, PR standards, and how to report bugs.

---

Documentation

Quickstart	Installation and first usage
API Reference	Full public API
Math Spec	Versioned update rule (nce_spec_v1)
Algorithm	Motivation and design
Player-aware mode	Per-example gradient stepping
Benchmarks	Reproducible results
Contributing	Development workflow
Changelog	Release history

---

License

Apache 2.0 — includes an explicit patent grant, which is the right default for optimization infrastructure.